Subscribe now

A Brief History

It started with pasting code into ChatGPT and hoping to get something useful back. Sometimes you hit gold; many times it made up some function that seemed like it should exist but ultimately didn’t. Then came GitHub Copilot, tab completion that was hit or miss. Then came the VSCode forks, Cursor and Windsurf, first with better tab completion, later introducing agents, the end state. Many more have since entered the fray, but the name of the game has been clear since then: lean on the models as much as possible.

Anyone who has consistently used generative AI coding tools over the last few years will tell you the same story. The end of 2025 was an inflection point. Before then, it was deeply suspicious if a model worked on a task for more than a few minutes. You could safely assume it’d gone down some strange rabbit hole, and you should probably rush to pull the plug before things got too messy. But then, in the last days of November, Anthropic released Claude Opus 4.5. That model, combined with the Claude Code harness, was a godsend. You could hand it large, fairly complex tasks. It would chug along, and a surprising amount of the time, it all just worked. Absolutely mind blowing, and a gut punch to developers who thought, maybe hoped, that AI coding might soon hit a wall. It seemed the labs were right. There may not be a wall.

The Right Bet

Even before Opus 4.5, developers loved Claude. Anthropic consistently, one release after the next, produced the best coding model, because that was their focus. They made the right bet: coding is the killer app. AI will undoubtedly reshape many industries, but programming is its first clear, dramatic win. It’s what people are willing to pay for and where the results are most visible. And not only is it a winner on the balance sheet, it also has a recursive benefit. A better coding model makes itself better tools.

OpenAI called a code red. At first they were slow off the mark. They released GPT 5.2-Codex the following December. Though some reported decent results, the model was too slow to be useful. But finally in early February, they pulled it off. GPT-5.3-Codex wasn’t just good, it was better. It handled complex, long-horizon tasks with more patience. Claude was too eager to start writing code. GPT-5.3-Codex took its time to gather the right context, and the code it produced was tighter. Though Claude still has an edge in frontend design and the general art of making things look nice, OpenAI has kept the lead. Many have soured on recent Claude releases, 4.6 and 4.7, while others sing the praises of GPT-5.4, the latest from OpenAI, released just a month after 5.3-Codex.

Compute

Anthropic has always been stingy with compute. The $20 Claude plan never got you as much as the equivalent from ChatGPT. But recently it’s been getting worse. They banned using Claude subscriptions with OpenClaw, and they’re testing removing Claude Code from their cheapest plan. OpenAI is doing just the opposite. They welcome OpenClaw, and their cheap plan keeps Codex. Every time Codex hits another million weekly users, and sometimes just because, they reset limits. Anthropic offers no such token jubilees. It can’t. All the outages and rationing make it obvious that they barely have the compute to handle current demand. Meanwhile OpenAI is wielding its compute advantage to keep existing users happy and entice new ones.

The Scientist and the Dealmaker

Early this year, Dario Amodei, CEO of Anthropic, went on the Dwarkesh Patel podcast and explained his reluctance to nab all the compute. Yes, revenue was scaling spectacularly well, but if growth came in even slightly below what he expected, and he’d bought too much compute, nothing in the world could stop him from going bankrupt. Reasonable, sensible answer. Exactly the kind of cautious, calculated response that makes me believe Dario has a pretty good relationship with his CFO.

Sam Altman, according to reporting, has a terrible relationship with his CFO, and it’s hard not to imagine that’s because he hasn’t been as restrained. He’s spent his time cutting every deal he could to get more compute, with seemingly little regard for the financials.

Only time will tell who comes out on top, or even survives. But right now this is one area where OpenAI’s recklessness was the right choice. Anthropic is running out of capacity to serve the users it already has, while OpenAI is raining down tokens from on high. Developers notice. They switch. And once they do, the recursive flywheel that made Claude the best coding model in the first place starts turning for the other team.

Subscribe now

Arcnem AI — Software development and AI consulting, based in Tokyo. Get in touch

Sure, these tools will give consumers pricing power—ditching a vendor for a homegrown solution is more feasible now than it was a year ago. But that has its own cost. If you make it, you have to maintain it, and who wants to do that? You’d rather have your employees focus on what actually generates revenue than build a second-rate calendar app you could’ve paid a few bucks for.

You pay for software to offload certain kinds of thinking to other teams. That’s always been the deal.

Every minute spent tweaking a boutique, internal project management system is time taken from perfecting the ad engine that pays the bills. The build cost went down. The opportunity cost barely moved.

So the vibe-coding-kills-SaaS panic feels overblown, but there’s another version of this story that’s harder to dismiss.

The fear isn’t that your customers replace you with a weekend project. It’s that the model companies eat the application layer entirely—that something like ChatGPT becomes the interface through which we do everything.

The model companies raised such enormous sums it was always obvious they’d come for this layer eventually. There’s a huge wave of applications building on top of these models, each claiming some trick, some moat that will keep it relevant, keep Claude from stealing their lunch. But does any of that plumbing actually matter? Richard Sutton’s “bitter lesson” suggests not. He posits that at lower levels of intelligence, you need orchestration—domain-specific tricks—to get the right output, but you obviate all of that by just having more intelligence. The scoreboard so far suggests he’s right. If he is, then everything built on top of the models becomes an increasingly thin layer. Why bother with a legal tech startup when a smarter Claude can handle contracts?

Aaron Levie, CEO of Box, makes the counterargument well: people love tools specific to their needs—a team that eats and sleeps your particular problem. That feels right to me. These companies should always have a place, but their position is unenviable. They’ve built castles on land they don’t own. The model companies can always beat them on price per token. And what if they start keeping the best models for themselves? Who survives the next leap in intelligence?

Every unlock gained by clever engineering is table stakes in the next model release, so the moat can’t be plumbing. It has to be something else. If you can’t name it, you don’t have it.

Subscribe now

Arcnem AI — Software development and AI consulting, based in Tokyo. Get in touch